Geomechanics Rock Property Uncertainty Workflow Using a Bayesian Machine Learning Framework

Abstract

ABSTRACT: Geomechanical rock properties have a broad impact on drilling, completion, production, and reservoir management decisions. Accurate quantification of measurements and correctly propagating uncertainty throughout the modeling process can improve decision quality and reduce the cost of field development. Laboratory rock property tests are generally sparse. To build a mechanical earth model for the entire field, empirical correlations are used to bridge the gap between lab data and petrophysical or geophysical measurements. Empirical correlations often rely on limited core test data from one or a few fields for a specific formation. This makes applying a correlation that is developed for a specific field to other fields difficult. A Bayesian machine learning approach is presented for modeling the rock properties. It provides a robust framework for building multivariate rock property models and evaluating uncertainties associated with the model. Core test data from many different fields with large variations of properties were used for this study. The workflow uses a hierarchical Bayesian regression, which enables the accurate learning of field-specific correlations in fields with much data and borrowing of correlation information from other fields in fields with little data. 1. INTRODUCTION Evaluating rock strength and elastic properties is fundamental to building a mechanical earth model (MEM). Most common geomechanical rock properties used in a MEM are unconfined compressive strength (UCS), Young’s Modulus (YM), Poisson’s ratio (PR), and friction angle (FA). A MEM has broad applications in drilling, completion, and production, such as calculating safe drilling windows, establishing injection limits for CO2 storage, enhanced oil recovery, and evaluating sand production risk, etc. (Plumb, 2000 and Ray et al 2007) Therefore, having the most accurate assessment of rock properties and the associated uncertainties is vital. Geomechanical rock properties are best characterized by direct laboratory measurements using downhole cores. However, coring can be a very expensive operation when factoring in equipment and rig costs. In most fields, cores are acquired in only a few carefully selected wells and intervals and, therefore, direct core measurements for a given field are limited. To build a MEM for the entire field, one must rely on empirical correlations using geophysical measurements, such as well logs and seismic data, and calibrate them to direct core measurements.